Wednesday Jul 8
FLOPSAnthropic

Claude Hides A 'Too Dumb' Tag

8JUL
MOCKEDSWAP

Anthropic's priciest model was quietly downgrading people without telling them. A developer found a tag called TOO_DUMB_TO_NEED_FABLE rerouting paid requests to a cheaper model. Backlash forced Anthropic to extend free access.

OpenCode developer Dax found the tag in Claude's logs this week. Fable 5 costs double Opus 4.8, yet the tag rerouted some requests elsewhere.

A Claude Code engineer's reply, 'I didn't expect you to look at the logs,' made it worse. Developers said they were paying premium prices for a cheaper model's answer. Anthropic extended free access to July 12.

The mechanism itself, a quality classifier, isn't unusual - hiding it is what stung. Trust in usage-based pricing takes one leaked log line to crack.

full brief & sources

⚡ Why this matters

  • Users can't budget for a model that might silently swap itself for a cheaper one.
  • The engineer's response turned a technical footnote into a trust story.
  • Anthropic reversing course under public pressure shows the backlash actually landed.

🔍 What happened

  • This week: OpenCode developer Dax found a 'TOO_DUMB_TO_NEED_FABLE' tag in Claude's system logs.
  • Fable 5 costs $10 per million input tokens, $50 per million output - double Opus 4.8.
  • Claude Code engineer Thariq Shihipar replied, 'Honestly, I didn't expect you to look at the logs.'
  • Developers said they were being billed premium rates for cheaper-model answers.
  • Jul 8: Anthropic extended free Fable 5 access on paid plans through July 12.

💬 Smart takes

  • Thariq Shihipar (Anthropic): 'Honestly, I didn't expect you to look at the logs.'
  • Developer community: paying double for a classifier that quietly downgrades your request defeats the point of choosing the expensive model.
  • Skeptic: routing to a cheaper model when the answer doesn't need the expensive one is a reasonable cost-saving feature, badly named and badly explained.

🧭 Where this goes

  1. LikelyAnthropic renames or publicly documents the routing classifier within weeks.
  2. PossibleFable 5 returns to full subscription inclusion once compute capacity catches up.
  3. Possibleother labs quietly check their own routing logic for similar undisclosed downgrades.
  4. Wild Cardthis becomes a case study in AI pricing transparency complaints to regulators.

🥄 The Spoon Take

The tech here isn't the problem - quietly downgrading a paid request without saying so is. Anthropic can defend the classifier as cost control, but the name it left in the logs undercut that. The backlash was about not being told.

🤔 Pushback

Anthropic reversed course within days, which is the system working, not a scandal - most vendors don't explain their routing logic at all.

·BigGo Finance - Tag discovery, backlash·Android Authority - Access extended·The New Stack - Five more days